- Munich Personal RePEc Archive

M PRA
Munich Personal RePEc Archive
Rainfall Drought Simulating Using
Stochastic SARIMA Models for Gadaref
Region, Sudan
Hisham Moahmed Hassan and Tariq Mahgoub Mohamed
University of Khartoum, Khartoum Technical College
2014
Online at http://mpra.ub.uni-muenchen.de/61153/
MPRA Paper No. 61153, posted 8. January 2015 13:38 UTC
Rainfall Drought Simulating Using
Stochastic SARIMA Models for
Gadaref Region, Sudan
Hisham Mohamed Hassan Alia
Tariq Mahgoub Mohamedb
a
Department of Econometrics and Social Statistics, University of Khartoum
E-mail: [email protected]
b
Khartoum Technical College
E-mail: [email protected]
ABSTRACT
Sudan is one of the countries which economy depends on rain fed agriculture and also facing
recurring cycles of natural drought. For many decades, recurrent drought, with intermittent
severe droughts, had become normal phenomenon in Sudan. This paper presents linear
stochastic models known as multiplicative seasonal autoregressive integrated moving
average model (SARIMA) used to simulate monthly rainfall in Gadaref station, Sudan. For
the analysis, monthly rainfall data for the years 1971–2010 were used. The seasonality
observed in ACF and PACF plots of monthly rainfall data was removed using first order
seasonal differencing prior to the development of the SARIMA model. Interestingly, the
SARIMA (0,0,5)x(1,0,1)12 model developed here was found to be most suitable for
simulating monthly rainfall over the Gadaref station. This model is considered appropriate
to forecast the monthly rainfall to assist decision makers establish priorities for water
demand, storage and distribution.
I. INTRODUCTION
Gadaref1 region lies in east central part of Sudan, at the border with Ethiopia. The mean
annual rainfall in this region, during the last four decades, is 617 mm. The annual number of
rainy days, (rainfall > 1 mm), is 111 days and the mean annual reference potential
evapotranspiration (ETO) using Penman/Monteith criterion for the region is about 2283mm
(Le Houérou, 2009). The region experiences very hot summer and temperature in the region
reaches up to 45o C in May. Generally the dry periods are accompanied with high
temperatures, which lead to higher evaporation affecting natural vegetation and the
agriculture of the region along with larger water resources sectors. Annual potential
evapotranspiration exceeds annual precipitation in this region. The rainfall exceeds
1
The word Gedarif (Gadaref, El Gadarif, Qadarif ) is derived from the Arabic phrase All
Gada-Ye-rif, meaning: he who had finished selling or buying should leave
1
evapotranspiration only in August and September (Etuk and Mohamed, 2014).The climate in
the Gedaref is semi-arid with mean annual temperature near 30o C (Elagib and Mansell,
2000).
A few researchers who have modeled rainfall drought using Standardized Precipitation
Index (SPI) as a drought indicator by SARIMA methods in recent times are Mishra and
Desai (2005) and Durdu (2010). For instance, Mishra and Desai (2005) fitted a SARIMA (1,
0, 0) x (1, 1, 1) 6 model to simulate and forecast SPI-6 in Kansabati river basin, India. Durdu
(2010) modeled the SPI-6 for Buyuk Menderes river basin located in the western part of
Turkey and fitted a SARIMA(1, 0, 0)x(2, 0, 1) 6 to it.
Time series model development consists of three stages identification, estimation, and
diagnostic checking (Box and Jenkins, 1970). The identification stage involves transforming
the data (if necessary) to improve the normality and stationary of the time series to
determine the general form of the model to be estimated. During the estimation stage the
model parameters are calculated. Finally, diagnostic test of the model is performed to reveal
possible model inadequacies to assist in the best model selection.
II.
METHODOLOGY
For more than half a century, Box–Jenkins ARIMA linear models have dominated many
areas of time series forecasting. Autoregressive (AR) models can be effectively coupled with
moving average (MA) models to form a general and useful class of time series models called
autoregressive moving average (ARMA) models. In ARMA model the current value of the
time series is expressed as a linear aggregate of p previous values and a weighted sum of q
previous deviations (original value minus fitted value of previous data) plus a random
parameter, however, they can be used when the data are stationary (Mishra and Desai 2005).
A time series is said to be stationary if it has constant mean and variance. This class of
models can be extended to non-stationary series by allowing differencing of data series.
These are called autoregressive integrated moving average (ARIMA) models.
Often time series possess a seasonal component that repeats every s observations. For
monthly observations s = 12 (12 in 1 year), for quarterly observations s = 4 (4 in 1 year). In
order to deal with seasonality, ARIMA processes have been generalized: The
full ARIMA model is called the SARIMA, a seasonal differencing element. The
regular ARIMA includes the AR polynomial and the MA polynomial the SARIMA model
incorporates both non-seasonal and seasonal factors in a multiplicative model. One
shorthand notation for the model is ARIMA(p, d, q) × (P, D, Q)S, with p = non-seasonal AR
order, d = non-seasonal differencing, q = non-seasonal MA order, P = seasonal AR
order, D = seasonal differencing, Q = seasonal MA order, and S = time span of repeating
seasonal pattern.
Without differencing operations, the model could be written more formally as
Φ(BS)φ(B)(xt - μ) = Θ(BS)θ(B)wt
(1)
The non-seasonal components are:
2
AR: φ(B) = 1 - φ1B - ... - φpBp
MA: θ(B) = 1 + θ1B + ... + θqBq
The seasonal components are:
Seasonal AR: Φ(BS) = 1 - Φ1BS - ... - ΦPBPS
Seasonal MA: Θ(BS) = 1 + Θ1BS + ... + ΘQBQS
Note that on the left side of equation (1) the seasonal and non-seasonal AR components
multiply each other, and on the right side of equation (1) the seasonal and non-seasonal MA
components multiply each other.
ARIMA (1, 0, 0) × (1, 0, 0)12
The model includes a non-seasonal AR(1) term, a seasonal AR(1) term, no differencing, no
MA terms and the seasonal period is S = 12, the non-seasonal AR(1) polynomial is φ(B) = 1
- φ1B. the seasonal AR(1) polynomial is Φ(B12) = 1 - Φ1B12.
The model is (1 - Φ1B12)(1 - φ1B)(xt - μ) = wt.
(2)
If we let zt = xt - μ (for simplicity), multiply the two AR components and push all but zt to the
right side we get
zt = φ1zt-1 + Φ1zt-12 + (-Φ1φ1)zt-13 + wt.
(3)
This is an AR model with predictors at lags 1, 12, and 13.
R can be used to determine and plot the PACF for this model, with φ1=.6 and Φ1=.5. That
PACF (partial autocorrelation function)
In this study, SARIMA models were used to simulate droughts based on the procedure of
models developments. The models are applied to simulate droughts using (SPI) series in
Gadaref region, (SPI-6 = SPI for 6 month).
After identifying models, it is needed to obtain efficient estimates of the parameters. These
parameters should satisfy two conditions namely stationary and invariability for
autoregressive and moving average models, respectively. The parameters should also be
tested whether they are statistically significant or not. The parameters values are associated
with standard errors of estimate and related t-values.
The drought events were calculated using the SPI. The data series from 1971 to 2010 were
used for model development for SPI-6 series.
There are two software packages which are used for time series analysis. These programs
are the SPSS 19 package and Eviews6.
3
IV. RESULTS AND DISCUSSION
Time series plot was conducted using the raw data, SPI 6_Gadaref, to assess its stability.
The assessments results are shown in fig. 1 it is clearly depicted that the time series are
stationary.
SPI6
FIG. 1: SPI 6 OF GADAREF STATION TIME SERIES
1971-2010
4
3
2
1
0
-1
-2
-3
-4
1975
1980
1985
1990
1995
2000
2005
2010
Stationary is also confirmed by the Augmented Dickey-Fuller Unit Root Test (ADF test) on
the data. The ADF test was conducted on the entire data. Table 2 shows ADF test results.
ADF test value -7.29548 less than critical vales -3.9778, -3.4194, -3.1323 all at 1%, 5%, and
10% respectively. This indicates that the series is stationary.
TABLE 2: ADF UNIT ROOT TEST (SPI6-GADAREF)
Station
Variable
ADF test
Gadaref
SPI_6
-7.29548
Level of
Confidence
Critical
Value
Probability
1%
-3.9778
0.0000
5%
-3.4194
0.0000
10%
-3.1323
0.0000
Result
stationary
In this step, the model that seems to represent the behaviour of the series is searched, by the
means of autocorrelation function (ACF) and partial auto correlation function (PACF), for
further investigation and parameter estimation. The behaviour of ACF and PACF is to see
whether the series is stationary or not.
For modelling by ACF and PACF methods, examination of values relative to auto regression
and moving average were made. An appropriate model for estimation of SPI_6 values for
the station was finally found.
4
Figure 2 shows the ACF and PACF, which have been estimated for SPI-6 for Gadaref
station. Many models for Gadaref stations, according to the ACF and PACF of the data,
were examined to determine the best model .The model that gives the minimum Akaike
Information Criterion (AIC) and Schwarz Criterion (SC) is selected as best fit model, as
shown in Table 3.
FIG. 2: ACF AND PACF PLOT FOR GADAREF STATION (SPI_6) SERIES
Autocorrelation
Partial Correlation
.|***** |
.|**** |
.|*** |
.|** |
.|* |
*|. |
*|. |
*|. |
*|. |
.|. |
.|. |
.|. |
.|. |
.|. |
.|. |
.|. |
.|. |
.|. |
.|. |
.|. |
*|. |
*|. |
*|. |
*|. |
.|***** |
.|. |
.|. |
.|. |
*|. |
**|. |
.|* |
.|. |
.|. |
.|. |
.|. |
.|. |
.|. |
.|. |
.|. |
.|. |
.|. |
.|. |
.|. |
*|. |
*|. |
.|. |
.|. |
.|. |
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
AC
PAC
Q-Stat
Prob
0.710
0.515
0.381
0.264
0.110
-0.087
-0.095
-0.074
-0.074
-0.055
-0.021
0.032
0.071
0.046
0.030
0.027
0.032
0.016
-0.014
-0.049
-0.089
-0.112
-0.144
-0.154
0.710
0.021
0.016
-0.036
-0.147
-0.235
0.168
0.065
-0.001
0.042
-0.004
-0.028
0.066
-0.063
-0.030
0.027
0.033
-0.011
0.002
-0.093
-0.088
-0.006
-0.035
0.000
240.58
367.28
436.84
470.20
476.05
479.71
484.06
486.67
489.31
490.78
490.99
491.49
493.93
494.98
495.42
495.77
496.29
496.42
496.51
497.72
501.65
507.92
518.29
530.14
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
TABLE 3: COMPARISON OF AIC FOR SELECTED MODELS, (SPI6_GADAREF)
Variable
SPI_6
Station
Model
AIC
Gadaref
ARIMA(1,0,0)
ARIMA(1,0,1)
ARIMA(2,0,1)
ARIMA(2,0,2)
ARIMA(0,0,1)
ARIMA(0,0,5)
SARIMA(0,0,5) (1,0,0)
SARIMA(0,0,5) (1,0,1)
SARIMA(0,0,5) (0,0,1)
SARIMA(1,0,0) (1,0,1)
2.104
2.108
2.113
2.102
2.335
2.010
2.036
1.999
2.104
2.092
5
The ACF and PACF correlograms, Fig. 2, and the coefficient are analyzed carefully and the
SARIMA model chosen is SARIMA (0,0,5) (1,0,1), as shown in table 3.
After the identification of the model using the AIC and SC criteria, estimation of parameters
was conducted. The values of the parameters are shown, in table 4. The result indicated that
the parameters are all significant since their p-values is smaller than 0.05 and should be used
in the model.
TABLE 4: SUMMARY OF PARAMETER ESTIMATES AND SELECTION CRITERIA (AIC),
(SPI6_GADAREF)
Variable
Coefficient
Std. Error
t-Statistic
Prob.
AR(12)
MA(1)
MA(2)
MA(3)
MA(4)
MA(5)
SMA(12)
-0.875450
0.697188
0.514251
0.408934
0.402012
0.394849
0.951476
0.025327
0.043122
0.050715
0.052794
0.050717
0.043373
0.013937
-34.56547
16.16768
10.14007
7.745840
7.926640
9.103536
68.27071
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
0.0000
R-squared
Adjusted R-squared
S.E. of regression
Sum squared resid
Log likelihood
Durbin-Watson stat
Inverted AR Roots
0.573186
0.567557
0.652726
193.8532
-454.9345
2.000692
.96+.26i
.26-.96i
-.70-.70i
.96+.26i
.47-.68i
-.26+.96i
-.70-.70i
Inverted MA Roots
Mean dependent var
S.D. dependent var
Akaike info criterion
Schwarz criterion
Hannan-Quinn criter.
.96-.26i
.26+.96i
-.70-.70i
.96-.26i
.47+.68i
-.26-.96i
-.70-.70i
.70+.70i
-.26-.96i
-.96+.26i
.70-.70i
.26-.96i
-.41-.73i
-.82
0.024048
0.992582
1.999716
2.062376
2.024386
.70-.70i
-.26+.96i
-.96-.26i
.70+.70i
.26+.96i
-.41+.73i
-.96-.26i
-.96+.26i
As considered in table 4 the model SARIMA (0,0,5) (1.0.1) has been selected as the one
with min AIC. The model has been identified and the parameters have been estimated. The
model verification is concerned with checking the residuals of the model to see if they
contain any systematic pattern which still can be removed to improve the chosen ARIMA.
All validation tests are carried out on the residual series. The tests are summarized briefly in
the following paragraph.
For a good model, the residuals left over after fitting the model should be white noise. This
is revealed through examining the autocorrelations and partial autocorrelations of the
residuals of various orders. For this purpose, the various correlations up to 24 lags have been
computed. The ACF and PACF of residuals of the model are shown in figure 3.
Most of the values of the RACF and RPACF lies within confidence limits except very few
individual correlations appear large compared with the confidence limits. The figure
indicates no significant correlation between residuals.
6
Autocorrelation
Partial Correlation
AC
PAC
Q-Stat
Prob
.|.
|
.|.
|
1
-0.001
-0.001
0.0001
.|.
|
.|.
|
2
-0.005
-0.005
0.0110
.|.
|
.|.
|
3
0.025
0.025
0.2931
.|.
|
.|.
|
4
-0.009
-0.009
0.3302
.|.
|
.|.
|
5
-0.012
-0.012
0.3967
.|.
|
.|.
|
6
-0.022
-0.023
0.6310
.|.
|
.|.
|
7
-0.039
-0.039
1.3437
.|.
|
.|.
|
8
0.018
0.018
1.4908
0.222
9
-0.076
-0.076
4.2551
0.119
*|.
|
*|.
|
.|.
|
.|.
|
10
-0.015
-0.014
4.3686
0.224
.|.
|
.|.
|
11
-0.022
-0.026
4.6047
0.330
.|.
|
.|.
|
12
-0.003
-0.001
4.6100
0.465
13
0.098
0.096
9.1690
0.164
.|*
|
.|*
|
.|.
|
.|.
|
14
0.017
0.015
9.3048
0.232
.|.
|
.|.
|
15
-0.046
-0.048
10.337
0.242
.|.
|
.|.
|
16
-0.013
-0.025
10.415
0.318
.|.
|
.|.
|
17
-0.005
-0.005
10.427
0.404
.|.
|
.|.
|
18
0.028
0.027
10.815
0.459
.|.
|
.|.
|
19
0.027
0.031
11.167
0.515
.|.
|
.|.
|
20
-0.039
-0.037
11.899
0.536
.|.
|
.|.
|
21
0.002
-0.006
11.900
0.614
.|.
|
.|.
|
22
-0.016
-0.009
12.022
0.677
.|.
|
.|.
|
23
-0.052
-0.044
13.317
0.649
.|.
|
.|.
|
24
0.044
0.045
14.266
0.648
FIG. 3: THE ACF AND PACF OF RESIDUALS FOR SPI-6 FOR GADAREF STATION
MODEL
The Ljung-Box Q-statistic is employed for checking independence of residual. From figure
3, ones can observe that the p-value is greater than 0.05 for all lags, which implies that the
white noise hypothesis is not rejected.
The Breusch-Godfrey Serial Correlation LM test accepts the hypothesis of no serial
correlation in the residuals, as shown in table 5. Durbin Watson statistic, (DW=1.999764),
also indicated that there is no serial correlation in the residuals.
7
Table No. 5: The Breusch-Godfrey Serial Correlation LM test (SPI6_Gadaref)
Breusch-Godfrey Serial Correlation LM Test:
F-statistic
0.030213
Prob. F(2,453)
0.9702
Obs*R-squared
0.001942
Prob. Chi-Square(2)
0.9990
Breusch-Godfrey Serial Correlation LM Test:
F-statistic
0.760940
Prob. F(12,443)
0.6909
Obs*R-squared
9.272109
Prob. Chi-Square(12)
0.6795
The Q-statistic and the LM test both indicated that the residuals are none correlated and the
model can be used. Since the coefficients of the residual plots of ACF and PACF are lying
within the confidence limits, the fit is good and the error obtained through this model is
tabulated in the table 6. The graph showing the observed and fitted values is shown in figure
(6.4).
TABLE NO. 6: ERRORS MEASURES OBTAINED FOR THE MODEL ARIMA (0,0,5) ,
(SPI6_GADAREF)
Error Measure
RMSE
MAE
R squared
Value
0.662
0.487
0.557
Figure 4 shows a very close agreement between the fitted model and the actual data.
Histogram of residuals for SPI_6 is shown in figure 5. This histogram shows that the
residuals are normally distributed. This signifies residuals to be white noise.
8
FIG.4: ACTUAL AND FITTED VALUES SARIMA (0, 0, 5)(1.0.1) , (GADAREF, SPI_6)
4
2
0
-2
3
2
-4
1
0
-1
-2
-3
1975
1980
1985
1990
Residual
1995
2000
Actual
2005
2010
Fitted
The graph of the (Q-Q) plot for the residual data look fairly linear, the normality
assumptions of the residuals hold, as shown in fig. 6.
90
Series: Residuals
Sample 1972M06 2010M11
Observations 462
80
70
60
50
40
30
Mean
Median
Maximum
Minimum
Std. Dev.
Skewness
Kurtosis
0.007362
0.004298
2.599692
-2.052816
0.648422
0.021672
3.808087
Jarque-Bera
Probability
12.60649
0.001830
20
10
0
-2
-1
0
1
2
FIG. 5: HISTOGRAMS OF RESIDUALS FOR SPI_6 GADAREF STATION
3
Quantiles of Normal
2
1
0
-1
-2
-3
-3
-2
-1
0
1
2
3
Quantiles of RESID
FIG. 6: (Q-Q) PLOT OF RESIDUALS FOR SPI_6 FOR GADAREF STATION
9
The K–S test is used to test the normality of residuals. It is observed that the Dcal is less than
Dtab at 5% significant level, shown in Table 6,( = 0.102 > 0.05).This test satisfies that the
residuals are normally distributed.
TABLE NO 6: K-S TEST CALCULATION OF RESIDUALS FOR SPI_6 SERIES, (GADAREF)
Kolmogorov-Smirnov Test
Most Extreme Differences (Dcal)
Dtable
0.057
0.063
One can note that all the model coefficients are statistically significant, each being more
than twice its standard error. The regression is very highly significant with a p-value of
0.0000, as high as 55.7% of the variation in data is accounted for by the fitted model. Figure
3 shows that the residuals are uncorrelated. Figure 4 shows a very close agreement between
the fitted model and the data. Therefore the fitted model is adequate. Fitted to the SPI_6 for
Gadaref station is the SARIMA (0, 0, 5) (1.0.1) model. Using various alternative arguments
it has been shown to be adequate.
III. Conclusion
In this paper linear stochastic model known as multiplicative seasonal autoregressive
integrated moving average model (SARIMA) was used to simulate droughts in Gadarif
region, Sudan. The models are applied to simulate droughts using Standardized Precipitation
Index (SPI) series, the results show that the fitted model is adequate to the SPI_6 for
Gadaref station is the SARIMA (0, 0, 5) (1.0.1) model.
IV. References
Abramowitz, M., Stegun, A. (eds.), (1965). Handbook of Mathematical Formulas, Graphs,
and Mathematical Tables. Dover Publications, Inc., New York, USA.
Box, G.E., Jenkins, G.M., (1970), Time Series Analysis; Forecasting and Control. HoldenDay series in time series analysis, San Francisco.
Durdu, O.F., (2010), Application of Linear Stochastic Models For Drought Forecasting in
The Buyuk Menderes River Basin, Western Turkey, Stochastic Environmental Research and
Risk Assessment, doi:10.1007/s00477-010-0366-3.
Elagib, N.A., M.G. Mansell., (2000), Recent Trends and Anomalies in Mean Seasonal and
Annual Temperature Over Sudan. Journal of Arid Environment 45(3): 263-288.
Etuk, E.H., Mohamed T.M., (2014). Time Series Analysis of Monthly Rainfall data for the
Gadaref rainfall station, Sudan, by SARIMA Methods. International Journal of Scientific
Research in Knowledge, 2(7), pp. 320-327.
10
Le Houérou, H.N. (2009): Bioclimatology and Biogeography of Africa. Verlag Berlin
Heidelberg, Germany.
Mishra, A. K., Desai, V. R., (2005), Drought Forecasting Using Stochastic Models.
Stochastic Environmental Research and Risk Assessment 19(5): 326-339.
Edwards, D.C., McKee, T.B., (1997). Characteristics of 20th Century Drought in the United
States at Multiple Scales. Atmospheric Science Paper 634.
Cacciamani, C., Morgillo, A., Marchesi, S., Pavan, V. (2007). Monitoring and Forecasting
Drought on a Regional Scale: Emilia-Romagna Region. Water Science and Technology
Library 62(1): 29-48
Thom, H.C.S., (1958). A Note on the Gamma Distribution. Monthly Weather Rev. 86, 117–
122.
Press WH, Flannery BP, Teukolsky SA, Vetterling WT. 1986. Numerical Recipes.
Cambridge University Press: Cambridge.
P. Angelidis, F. Maris, N. Kotsovinos and V. Hrissanthou, (2012). Computation of Drought
Index SPI with Alternative Distribution Functions”, Water Resour ManageVol. 26 p2453 –
2473.
McKee, T. B.; N. J. Doesken; and J. Kleist. (1993). “The relation of drought frequency and
duration to time scales.” Proceedings of the Eighth Conference on AppliedClimatology; pp.
179–84. American Meteorological Society, Boston.
Tsakiris, G., Vangelis, H. (2004). Towards a Drought Watch System based on spatial SPI.
Water Resources Management 18 (1): 1-12.
Mishra, A. K., Desai, V. R., (2005), Drought Forecasting Using Stochastic Models.
Stochastic Environmental Research and Risk Assessment 19(5): 326-339.
Durdu, O.F., (2010), Application of Linear Stochastic Models For Drought Forecasting in
The Buyuk Menderes River Basin, Western Turkey, Stochastic Environmental Research and
Risk Assessment, doi:10.1007/s00477-010-0366-3.
11